Iterative Neural Network Search for Image Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image matching methods require evaluating a large number of windows or pixels, making the search process inefficient and less robust to background clutter, occlusions, and geometric transformations.
Innovation Solution
A search apparatus and method that utilize multiple trained neural networks to iteratively refine the search path for a target region within media data, reducing the number of candidate regions to be evaluated.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If exhaustive search with sliding windows is used, then matching accuracy is satisfactory, but the number of windows to be evaluated becomes huge, making the search process prohibitive
Solution Approach 1:
The patent divides the exhaustive search process into multiple stages: initial candidate generation, candidate filtering, and final verification. By segmenting the search space and applying different evaluation strategies at each stage, the method maintains matching accuracy while dramatically reducing the total number of windows evaluated.
Solution Approach 2:
The patent performs preliminary actions by generating initial candidate regions using approximate matching criteria before conducting detailed evaluation. This preliminary filtering eliminates obviously mismatched regions early in the process, reducing the computational burden of exhaustive search while preserving accurate matches.
2Productivity
If pruning is introduced to reduce run time by skipping unnecessary windows, then search speed improves, but robustness to background clutter and occlusions deteriorates
Solution Approach 1:
The patent incorporates feedback mechanisms where evaluation results from candidate regions are used to refine subsequent search decisions. The system learns from positive and negative examples during evaluation, adjusting its pruning strategy to maintain robustness against background clutter while preserving search speed improvements.
Solution Approach 2:
The patent dynamically adjusts evaluation parameters and thresholds based on the complexity of the search scenario. By changing parameters such as matching thresholds and candidate selection criteria adaptively, the system maintains robustness to occlusions and background clutter while optimizing search speed for different conditions.
3Productivity
If a small subset of pixel pairs is used for matching, then computation cost is reduced and matching is accelerated, but accuracy in complex scenarios decreases
Solution Approach 1:
The patent transitions from evaluating individual pixel pairs to evaluating regions and patterns formed by multiple pixels. By changing the dimensionality of the matching unit from 1D pixel pairs to 2D/3D spatial patterns, the system achieves both computational efficiency and high accuracy in complex scenarios with background clutter and geometric transformations.
Data Source
AI summary
A search apparatus for searching media data for a target region that matches query data includes a first feature extraction unit configured to extract a first feature vector from the query data using a first trained neural network; a second feature extraction unit configured to obtain a first region from the media data and extract a second feature vector from the first region using a second trained neural network; a localization unit configured to determine a candidate for the target region using a third trained neural network, based on the first feature vector, the second feature vector, and the first region or a location of the first region; and a control unit configured to repeat the operations of the second feature extraction unit and the localization unit until a predetermined condition is satisfied, by using the determined candidate for the target region as the first region to be used by the second feature extraction unit.